Dear Aventine Readers,

The next time you get a check-up, there’s an excellent chance that an AI scribe, and not your doctor, will be recording the conversation and writing up the clinical notes. So far, physicians report that AI scribes relieve stress, and patients are pleased to have their doctor’s complete attention. But the big test for adoption is whether AI scribes will save money, and that picture isn’t nearly so clear.

Thanks for reading,

Danielle Mattoon 
Executive Director, Aventine

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

When Shreya Shah, a primary care physician at Stanford Primary Care Portola Valley in San Mateo County, California, starts a consultation, she fires up DAX Copilot, an AI scribe built by Microsoft, to take notes. The software records the conversation, uses large language models to make sense of what was said, then writes up notes of the consultation that the physician checks before they’re added to the patient’s electronic health record.

Shah doesn’t think the technology is saving her much time. But she does think the quality of care she’s delivering is improving. She spends less time looking at her computer and more time eye-to-eye with patients, talking through her medical reasoning in more detail than she used to. She’s also able to ensure her instructions and orders are properly written up and printed so that the patients can take them home. “I think I'm giving patients probably what they wanted to hear, and more information,” she said.

Her experience is becoming increasingly commonplace. AI scribes were one of the first examples used to demonstrate how AI could revolutionize healthcare. Since their introduction around the turn of the decade, adoption has soared. A March 2026 survey by the American Medical Association found that 80 percent of physicians now use some form of AI in a professional context, double the percentage in 2023. AI scribes in particular have gotten significant traction in the past two years, with 40 percent of respondents saying they now used them, up from just 14 percent in 2024. 

But AI scribes are revealing a surprising adoption dynamic. They have been eagerly embraced, credited with significant reduction in physician burnout and are popular with users. But contrary to what early boosters promised, they have not led to increased patient loads or immediate cost savings. This presents healthcare — and potentially other industries — with a provocative and potentially revealing question: How to weigh investment in a technology that doesn’t directly cut costs or obviously increase efficiency, but instead improves levels of service, reduces strain on employees and in doing so only indirectly boosts revenues?

AI in the clinic 

Rebecca Mishuris is a primary care physician and Chief Health Information Officer at Mass General Brigham who oversaw the rollout of AI scribes across the hospital group and immediately discovered that  doctors are desperate for them. "Never in my career, and probably never again, will there be a technology that [doctors] are clamoring for [as much]," she said. Mass General Brigham reported to the Peterson Health Technology Institute that approximately 90 percent of its ambulatory primary care physicians requested access to scribes.

Yes, scribes raise concerns about issues like bias, privacy and complacency but surveys have shown support for the technology among physicians and patients. The American Medical Association survey found that 76 percent of physicians think that AI will have a positive impact on patient care, up from 65 percent in 2023. In a 2025 Permanente Medical Group study of AI scribe use by 7,260 physicians across more than 2.5 million patient encounters, 47 percent of patients said their doctor spent less time looking at their computer, and 39 percent said their doctor spent more time speaking directly to them.

Despite not seeing obvious efficiency gains, physicians report other benefits, including lower rates of burnout. One observational study, of 1,430 physicians across Mass General Brigham and Emory Healthcare, found burnout rates fell by roughly 20 percentage points after the introduction of an AI scribe, a result that was sustained a year later. Those reductions are "unheard of in the burnout literature," said Mishuris, who helped run the study. "There's no other intervention that decreases burnout like that, technology or otherwise."

Less clear is what causes that reduction. Measurements of time saved vary drastically. Some AI scribe companies claim that doctors are saving up to two hours per day. Research studies suggest far less. In a five-center study across 8,581 clinicians that included five large teaching hospitals, scribes saved physicians 16 minutes of documentation time for every eight hours of patient care. Researchers increasingly think the mechanism is not minutes saved but cognitive load lifted: less multitasking during consultations, fewer details to hold in working memory, no need to type while talking. In a Stanford study looking at how physicians experience AI scribes co-authored by Shah, one doctor told the team: "I don't know if it saved time, but it saved anxiety." 

The lack of straightforward efficiency gains raises an important question about how willing healthcare managers might be to pay for the tools. If AI scribes don't free up doctors to see more patients, what is the justification for a tool that costs between $200 and $600 per doctor per month? Research published in January suggests the tools may pay for themselves regardless, at least at the lower end. A study at UCSF of over a million patient encounters found that physicians using AI scribes saw an increase in their relative value units — an industry measure of the complexity and expertise that go into a patient visit that helps determine reimbursement — equivalent to about $3,044 annually per doctor. At the cheaper end of the market at least, that more than covers the extra cost.

"There's a saying that clinicians don't get paid for the care that they provide, they get paid for the care that they documented they provide," said Matt Troup, a clinical strategy principal at the AI scribe company Abridge. AI scribes, he argued, capture more of the complexity of a consultation and keep more accurate records. "I think that is a rising tide for all of healthcare," Troup said.

In a recent article, Shah argued there is another reason hospitals and clinics should embrace AI. Burnout can as much as triple the likelihood of physician turnover, and replacing each departing physician can cost up to $1 million. The indirect savings of AI adoption, she argues, are significant.

What comes next

Current use of AI scribes may be just the tip of the iceberg. Far broader adoption is expected in coming months, largely because the healthcare software giant Epic Systems recently announced it would roll out its own AI scribe system. As AI use in primary care continues to grow, agentic systems are expected to help manage doctors' inboxes, handle prior authorization paperwork and process refills, among other tasks. The doctors Aventine spoke to think adoption of these newer tools could have a greater and more immediate impact on patients than note-taking. 

Yet there are concerns about whether future adoption will be equally accessible to all doctors. A survey of 2,174 US hospitals carried out by researchers from the US Department of Health and Human Services and the University of Minnesota School of Public Health, published in December 2025, found that “hospitals in multihospital systems were more than twice as likely to be early adopters of generative AI or fast followers than independent hospitals … and rural hospitals were substantially less likely to be early adopters or fast followers than their counterparts.” This is “the continuation of a digital divide” seen historically in healthcare, the survey authors wrote. 

Mishuris is hopeful that the situation might not play out that way. The AI scribe market is highly competitive, she argues, and cheaper options are already available to providers who cannot afford the most expensive enterprise offerings. 

Of course, providers aren’t alone in using AI to help manage healthcare: Insurance firms are building tools to analyze and approve — or, perhaps more concerning, deny— claims. Researchers from Stanford have outlined how, in the near future, generative AI used by insurers could predict a patient's care needs, potentially incorrectly, in ways that even the developers cannot fully explain, leading to outcomes that are hard to challenge. "It begs the question of … if two computers are fighting about this, should we re-examine the underlying policy?" said Mishuris. That, however, is a situation with no immediate solution in sight.

Doctors, in any case, feel better able to do their jobs. "I look my patients in the eye now and I am not tapping away at the keyboard during the visit," said Mishuris. "It used to be that I would get home and my husband would know if I had a really long day in the clinic. Now I leave the clinic with my notes done and with a clear mind."

AI is really starting to advance mathematics. An OpenAI model has disproved a mathematical conjecture that remained unresolved for nearly 80 years, a result some mathematicians are calling the most significant AI achievement in pure mathematics so far. The problem, posed by the mathematician Paul Erdős in 1946, is deceptively simple: Given a collection of points on a plane, how many pairs of points can be exactly the same distance apart? Erdős believed the optimal arrangement resembled square grids, implying the number of equal-distance pairs could grow only modestly as the number of points increased. No one managed to prove or disprove that intuition for decades.The OpenAI system found a counterexample by borrowing techniques from algebraic number theory, solving the problem in a higher-dimensional space before mapping the result into two dimensions. Human mathematicians have verified the proof and researchers at Anthropic claim to have replicated it. Tim Gowers, a Fields Medal-winning mathematician from the University of Cambridge who helped validate the result, described it as “a milestone in AI mathematics.” He added that AI isn't yet overtaking human mathematicians, but showing that it’s well-suited to specific problems where answers can be found by combining knowledge from disparate fields, which humans struggle with. 

What if your home became an AI data center? That’s the idea being explored by Span, a startup known for smart circuit breakers, which is testing air conditioner-size AI computing units installed outside people’s homes. As Scientific American reports, each box contains 16 Nvidia GPUs and can consume 12.5 kilowatts of power. That’s a lot of electricity: One unit operating continuously would use roughly as much energy in three days as the average US household consumes in an entire month. But homeowners won’t be paying. If they have a unit, they would receive discounted electricity and internet, plus compensation tied to how much compute their hardware provided. Span’s idea is to create a distributed data center, with small inference nodes across neighborhoods that tap spare capacity already available on the grid. Around 8,000 such units could provide computing power comparable to a medium-size 100-megawatt data center, but for now Span plans to launch a 100-home pilot in the southwestern US later this year, integrated into new housing developments. Together, those homes would provide about 1.2 megawatts of computing power. Skeptics have argued that coordinating workloads across thousands of tiny nodes could prove expensive, and utilities may oppose the idea of giving over spare grid capacity to AI infrastructure, particularly as electric vehicles and electrified heating systems place increasing strain on local networks. Still, the concept hints at a future of ever more distributed AI compute. 

Underground rock could produce clean hydrogen and lock away CO₂. Researchers at the University of Texas at Austin have demonstrated a process that, in theory, could generate clean hydrogen fuel, permanently sequester carbon dioxide and even produce geothermal energy. The chemistry relies on a natural geological process called serpentinization. The idea, New Scientist reports, is to turbocharge it by pumping water enriched with CO₂ into iron-rich volcanic rock deep underground. As the water reacts with the rock, hydrogen is released. At the same time, carbon dioxide mineralizes into solid carbonate minerals, trapping it underground. Laboratory experiments suggest that the process works, and researchers found that adding a nickel chloride catalyst significantly boosts hydrogen production. Finally,  because the injected water would heat up underground, the process could also be used to create  geothermal energy. The major challenge is efficiency: Right now, the approach captures only around 0.5 percent of the theoretically available hydrogen, and researchers estimate commercial viability may require at least doubling that. One possible solution is drilling deeper, where higher temperatures accelerate the reactions. Startups including Vema Hydrogen, GeoRedox Corporation and Anning Corporation are all exploring related approaches. The difficulty will be turning such an impressive concept into a commercial process based on messy real-world geology. But if it works, suitable iron-rich rocks are available globally, and researchers estimate that modestly efficient systems could generate volumes of clean hydrogen that collectively far exceed current global production.

Meta goes big on the Bayou, from Bloomberg Businessweek, and Building an AI Data Center in Pine Island, Minnesota, from The Paris Review
5,000 words, or about 21 minutes, and 2,800 words, or about 11 minutes

Here are two stories that explore the same theme: what happens when gigantic AI infrastructure projects arrive in small American communities. The Bloomberg Businessweek story focuses on Meta’s Project Hyperion in Richland Parish, Louisiana, a development that could eventually become the world’s largest AI data center, consuming up to 5 gigawatts of power, roughly comparable to the average daily electricity demand of New York City. Meta reportedly used shell companies, NDAs and legislative maneuvering to keep details secret while assembling land and financing. The consequences are already rippling through one of America’s poorest regions: Local professionals like nurses and police officers are leaving jobs for better-paying opportunities on the project, property values are climbing rapidly and local schools may need to relocate. The Paris Review piece, meanwhile, is more atmospheric, following Google’s proposed 482-acre Project Skyway facility in Pine Island, Minnesota, a small town better known for its cheese festival. The story captures a community processing the arrival of an industrial project so large it feels almost unreal with a mix of anger, anxiety and resignation. Together, the accounts expose an often unexamined byproduct of the AI-boom: the disruption of local communities when data centers come to town. 

Are Prediction Markets Good for Anything? from Asterisk
4,000 words, or about 16 minutes

CEOs of prediction platforms like Polymarket and Kalshi argue that they are creating markets that can generate accurate forecasts about the future, not gambling websites. This essay by forecasting expert Dan Schwarz, who previously built an internal prediction market at Google, explores whether they can deliver anything close to what they are promising. He evaluates prediction markets across five potential benefits: monitoring risk, interpreting breaking news, forecasting policy outcomes, holding leaders accountable for claims and surfacing genuinely new information. His findings are mixed. There are areas in which the markets appear genuinely useful: Geopolitical conflict markets, for instance, can provide valuable signals for traders, businesses, governments and journalists. But even there, the markets are often better at quantifying people’s existing anxieties than providing new insight. At other tasks, prediction markets are worse. Markets tied to emerging technologies or to broad political events like conflicts, for instance, often generate noisy or low-quality information. Schwarz argues that most activity on these platforms is not really about collective intelligence at all: It is sports betting, election spectacle and crypto-adjacent speculation dressed up in the language of information markets. 

What it takes to get a job at Anthropic, from Bloomberg Businessweek
1,900 words, or about 8 minutes

Getting hired at one of the world’s fastest-growing AI labs sounds a bit like applying to an elite university, interviewing for a hedge fund and attending therapy all at once. Anthropic offers life-changing compensation packages and the prospect of enormous payouts when the company goes public, which it filed to do earlier this week (after this article was published). Some senior executives have reportedly accepted far more junior titles to get a foot in the door. The hiring process itself sounds rigorous but not especially unusual at first blush: multiple interviews, technical assessments, extensive evaluations of skills. (It is worth noting that, ironically, candidates are discouraged from using AI tools during the process.) What appears to set Anthropic apart, though, is its obsession with culture. According to this piece, CEO Dario Amodei says he spends as much as 40 percent of his time maintaining the company’s culture, and Anthropic places extraordinary weight on determining whether candidates align with its values and worldview. Candidates are asked probing questions about their motivations, decision-making and beliefs. Some candidates describe it as invasive, likening it more to a therapy session than a job interview. The process is unusual enough — and the rewards large enough — that a cottage industry of interview coaches has emerged to help applicants navigate it, and candidates routinely spend thousands of dollars in preparation. But hey, that’s probably worth it to land a job that could provide life-changing compensation.